Si

XDa

b

c

a

ARRA

KMIRSGC

1

rdobap

isitl

0d

Journal of Neuroscience Methods 199 (2011) 129– 139

Contents lists available at ScienceDirect

Journal of Neuroscience Methods

jou rna l h om epa ge: www.elsev ier .com/ locate / jneumeth

ocial network theory applied to resting-state fMRI connectivity data in thedentification of epilepsy networks with iterative feature selection

iaohui Zhanga,∗, Fuyuze Tokoglua, Michiro Negishia, Jagriti Aroraa, Scott Winstanleyc,ennis D. Spencerc, R. Todd Constablea,b,c

Department of Diagnostic Radiology, Yale University School of Medicine, TAC N128, New Haven, CT 06511, USADepartment of Biomedical Engineering, Yale University School of Medicine, New Haven, USADepartment of Neurosurgery, Yale University School of Medicine, New Haven, USA

r t i c l e i n f o

rticle history:eceived 24 November 2010eceived in revised form 13 April 2011ccepted 14 April 2011

eywords:edial temporal lobe epilepsy (MTLE)

nterictalesting state fMRIocial network theoryraph theorylassification

a b s t r a c t

Epilepsy is a brain disorder usually associated with abnormal cortical and/or subcortical functional net-works. Exploration of the abnormal network properties and localization of the brain regions involved inhuman epilepsy networks are critical for both the understanding of the epilepsy networks and planningtherapeutic strategies. Currently, most localization of seizure networks come from ictal EEG observa-tions. Functional MRI provides high spatial resolution together with more complete anatomical coveragecompared with EEG and may have advantages if it can be used to identify the network(s) associatedwith seizure onset and propagation. Epilepsy networks are believed to be present with detectable abnor-mal signatures even during the interictal state. In this study, epilepsy networks were investigated usingresting-state fMRI acquired with the subjects in the interictal state. We tested the hypothesis that socialnetwork theory applied to resting-state fMRI data could reveal abnormal network properties at the grouplevel. Using network data as input to a classification algorithm allowed separation of medial temporallobe epilepsy (MTLE) patients from normal control subjects indicating the potential value of such net-

work analyses in epilepsy. Five local network properties obtained from 36 anatomically defined ROIswere input as features to the classifier. An iterative feature selection strategy based on the classificationefficiency that can avoid ‘over-fitting’ is proposed to further improve the classification accuracy. An aver-age sensitivity of 77.2% and specificity of 83.86% were achieved via ‘leave one out’ cross validation. Thisfinding of significantly abnormal network properties in group level data confirmed our initial hypothesisand provides motivation for further investigation of the epilepsy process at the network level.

. Introduction

Epilepsy is a common brain disorder characterized by recur-ent unprovoked seizures, associated with functionally abnormal,istributed cortical and subcortical network(s). The investigationf abnormal functional network properties and localization of therain regions responsible for seizure generation and propagationre key issues for both understanding the epilepsy networks andotentially guiding therapeutic strategies.

The evidence for human epilepsy networks comes mostly fromctal and pre-ictal EEG observations. In order to understand thepatiotemporal dynamics of seizure activity, measures either from

ndividual electrodes, for example power spectrum, or from elec-rode pairs such as correlation, coherence and synchronizationikelihood have been applied (Bertashius, 1991; Ponten et al.,

∗ Corresponding author. Tel.: +1 203 737 5995; fax: +1 203 785 6643.E-mail address: xiaohui.zhang@yale.edu (X. Zhang).

165-0270/$ – see front matter © 2011 Elsevier B.V. All rights reserved.oi:10.1016/j.jneumeth.2011.04.020

© 2011 Elsevier B.V. All rights reserved.

2007). In human medial temporal lobe epilepsy (MTLE), intracra-nial EEG (icEEG) has shown evidence of a medial temporal limbicnetwork which is a bilateral, cortical/subcortical network thatincludes the hippocampal formation, amygdala, entorhinal cortex,medial thalamus, and inferior frontal lobe (Spencer, 2002). Basedon human icEEG recordings of spontaneous seizures, the entirenetwork may variably participate in the genesis and expressionof seizure activity. The initial electrical events are expressed invarious network locations, and may vary from seizure to seizurein an individual (Spencer et al., 1992; Spencer and Spencer,1994; Spencer, 2002; Wennberg et al., 2002). Locational andmorphologic variability in apparent “seizure onset” is typical ofspontaneous seizures recorded within the human brain in MTLE,but is difficult to recognize with the sparse sampling providedby intracranial recording. Spencer and Spencer (1994) and So

(1991) further observed that seizure onset can be synchronousover multiple areas, with involvement of extrahippocampal andsubcortical regions. Global changes in epilepsy networks havebeen observed in several states including interictal, before rapid

1 scienc

dsrwl2

ociwmTcb(ttttrsaembiifs

rtjorivnpswcBemsmgwdasiaiLfiptw2

ip

such local network measures can differentiate a patient’s epilepsy

30 X. Zhang et al. / Journal of Neuro

ischarge, during rapid discharge (early ictal), during seizurepreading (late ictal) and postictally (Ponten et al., 2007). Tempo-al lobe functional connectivity was found to be lower in patientsith longer TLE history while longer TLE duration was corre-

ated with more random network configuration (Dellen et al.,009).

Although icEEG has good temporal resolution, its low spatial res-lution and the invasive nature of the procedure especially in deeportical layers limits its widespread utilization. Compared withcEEG, functional MRI provides higher spatial resolution together

ith whole-brain anatomical coverage. Functional connectivityapping using fMRI was first described by Biswal et al. (1995).

his approach utilizes the blood oxygenation level dependentontrast (BOLD) signal and measures the temporal correlationsetween different brain regions in a single subject over timeBiswal et al., 1995; Hampson et al., 2002; Lowe et al., 1998)ypically during rest. Low frequency (<0.1 Hz) spontaneous fluc-uations lead to temporal changes in the BOLD signal and theseemporal changes are often highly correlated across brain regionshat participate in similar networks. There is also evidence thatesting-state connectivity can be related to behavioral variablesuch as task performance (Hampson et al., 2006a,b) and it isltered in several clinical populations (Buckner et al., 2009; Mullent al., 2010). Since epilepsy networks are persistently abnor-al, we test the hypothesis that such networks could be defined

y the extent and strength of their components in the inter-ctal state. To detect abnormalities we compare the networksdentified in epilepsy patients in the interictal state with theunctional networks of a large set of control subjects in thistudy.

Specifically we investigate the extent to which we can extractepresentative features to classify functional connectivity networkshereby differentiating MTLE patients from healthy control sub-ects. Social network analysis was adapted to extract informationn the functional networks in the brain from the high-dimensionalesting-state fMRI data. Social network analysis has emerged as anmportant technique in sociology and it plays a significant role inarious fields such as economics, biology, and psychology. Socialetwork analysis provides tools that can be used to interpret com-licated coupling topologies in brain networks. Recent studiesuggest that networks derived from brain activity possess a “small-orld” topology characterized by dense local clustering and most

onnections have short path length (Sporns et al., 2004; Basset andullmore, 2006; Ponten et al., 2007; Stam et al., 2007). Bucknert al. (2009) investigated intrinsic cortical hubs via the degreeap, and the functional connectivity networks associated with

uch hubs. Such analysis methods have been applied in epilepsyost recently with Kramer et al. (2008) examining the emer-

ent network topology found in electrocorticography data. In theirork, six quantitative measures including average path length,egree, closeness, clustering coefficient, betweenness centralitynd betweenness centralization were used to identify statisticallyignificant changes in network topology between ictal and pre-ctal states. Global network properties such as average path lengthnd average clustering coefficient have also been studied in inter-ctal epilepsy networks via EEG/MEG (Horstmann et al., 2010).iao et al. (2010) investigated the interictal epilepsy network withMRI using network properties including degree, n-to-1 connectiv-ty, clustering coefficient, shortest path lengths and small-worldroperties. Recent work has demonstrated increasing evidencehat both cognitive and psychiatric disturbances are correlatedith functional network architectural features (Reijneveld et al.,

007).Unlike most of previous studies that focused on the ictal state,

n this work, by using social network theory, abnormal networkroperties were studied at both the individual and group level

e Methods 199 (2011) 129– 139

resting-state fMRI data. At the individual subject level, the net-work metrics associated with various regions of interest (ROIs)were found in both healthy control subjects and epilepsy patients.Group level, analysis was then performed in an attempt to deter-mine if a classification strategy based on network properties coulddistinguish medial temporal lobe epilepsy (MTLE) patients from thehealthy control subjects. This analysis also allowed us to extractthe network features that provided the best classification of thetwo groups of subjects. Five local network properties from 36ROIs were entered as features to the classifier. To further improvethe classification accuracy and at the same time avoid ‘over fit-ting’, a feature selection strategy was proposed. The classificationaccuracy was evaluated via ‘leave one out’ cross validation. Thenetwork properties of various ROIs were ranked based on their abil-ity to separate patients from healthy controls. The abnormal ROIsfound both in individual patient and group level compared withhealthy control subjects can aid in the understanding of epilepsynetworks and potentially provide guidance in choosing therapeuticoptions.

2. Theory

A network G consists of a set of vertices and a set of edges. Eachedge links two vertices with a value defined as weight. In this paper,a weighted undirected network is adopted.

Five network properties including: Degree, Strength, Closeness,Clustering coefficient and Betweenness centrality (Wasserman andFaust, 1994) were selected for brain network analysis. The defini-tions of these properties follow.

Degree: Let eij denote the connection between vertex i and vertexj. If i and j are connected, eij = 1, otherwise eij = 0. The degree D(i) ofa vertex i is the number of vertices it connects to, D(i) =

∑jeij.

Strength: Let wij denote the weights between vertex i and vertexj. The strength S(i) of a vertex i is the sum of all connection weights,S(i) =

∑jwij .

Closeness: The closeness of vertex i is defined as the number ofvertexes reachable from i divided by the summed distance to thesereachable vertexes. The distance between two vertices is definedas the inverse of the weight. C(i) = g − 1/(

∑gj=1d(i, j)), g is the con-

nected group size of vertex i, d(i,j) is the distance between vertex iand j, j /= i.

Clustering coefficient: The clustering coefficient � measures thetightness of a connection in a local sub-network. Let N(i) denotethe set of vertices that connect to vertex i and S the numberof vertices in N(i). For an undirected network, the total numberof edges within N(i) is given by (1/2)

∑j∈N(i)

∑k∈N(i)ejk. If every

point in N(i) is connected to every other point in the set, the totalnumber of edges is (1/2)S(S − 1). The clustering coefficient is theratio between the actual number of edges and the total number ofpossible edges, �(i) =

∑j∈N(i)

∑k∈N(i)ejk/(S(S − 1)). As for weighted

network, the generalization proposed by Onnela et al. (2005)was adopted in this work. In this case, �(i) =

∑jk(wjiwikwjk)1/3/

(S(S − 1))If the vertices are well connected locally, � will be close 1.Betweenness centrality: The betweenness centrality �(i) is the

number of shortest paths between any two vertices that travelthrough i. It is usually normalized by dividing its maximumvalue.

These five network metrics reveal the specific local networkproperties for each vertex. Our goal in this work is to determine if

network from a healthy control network and hence locate theabnormal brain regions. These properties were calculated usingthe software package of Pajek (Nooy et al., 2005) and the BrainConnectivity Toolbox (Sporns et al.).

science Methods 199 (2011) 129– 139 131

3

3

f3

afs3aewiFtdu

3

lctuwm

3

blaflp2

isptbwicc

3

ttv

A

B

ex

Gen

der

Age

Firs

t

un

pro

voke

dse

izu

reD

ura

tion

ofM

TLE

Seiz

ure

area

Path

olog

y

Seiz

ure

s

freq

uen

cy

M

11

3

8

Left

ante

rom

edia

l tem

por

al

lobe

and

hip

poc

amp

us

Neo

cort

ex

wit

h

slig

ht

non

-sp

ecifi

c

abn

orm

alit

ies

1–2/

mon

thF

43

40

3

Rig

ht

ante

rom

edia

l tem

por

al

lobe

and

amyg

dal

aan

d

pes

tum

orC

han

ges

con

sist

ent

wit

h

amyg

dal

e

dys

pla

sia

2–4/

day

M

38

29

9

Left

ante

rom

edia

l tem

por

al

lobe

and

hip

poc

amp

us

Hip

poc

amp

al

scle

rosi

s

2/m

onth

M52

44

8

Left

amyg

dal

e

and

hip

poc

amp

us

Hip

poc

amp

us

wit

h

neu

ron

al

loss

and

glio

sis,

con

sist

ent

wit

h

hip

poc

amp

al

scle

rosi

s1–

3/w

eek

F

46

5

41

Left

ante

rom

edia

l tem

por

al

lobe

and

hip

poc

amp

us

Hip

poc

amp

al

scle

rosi

s

1/2

mon

ths

F

51

2.5

48.5

Left

ante

rior

med

ial t

emp

oral

Hip

poc

amp

us

wit

h

neu

ron

al

loss

and

glio

sis,

con

sist

ent

wit

h

hip

poc

amp

al

scle

rosi

s3–

5/d

ay

F

40

13

27

Left

ante

rom

edia

l tem

por

al, a

myg

dal

a

and

hip

poc

amp

us

Neu

ron

al

loss

and

glio

sis,

con

sist

ent

wit

hh

ipp

ocam

pal

scle

rosi

s6–

7/w

eek

F

27

19

8

Rig

ht

ante

rom

edia

l tem

por

al

lobe

, am

ygd

ale,

and

hip

poc

amp

us

Hip

poc

amp

us

wit

h

neu

ron

al

loss

and

glio

sis,

con

sist

ent

wit

h

hip

poc

amp

al

scle

rosi

s4–

5/m

onth

F40

20

20

An

teri

or

med

ial t

emp

oral

and

hip

poc

amp

us

Hip

poc

amp

al

scle

rosi

s

6–10

/mon

thM

35

17

18

Tem

por

al

lobe

Abn

orm

al

hip

poc

amp

us

4–8/

mon

thF

18

9

9

Left

med

ial t

emp

oral

lobe

Hip

poc

amp

al

atro

ph

y

4–6/

mon

thF

50

5

45

Rig

ht

ante

rom

edia

l tem

por

al

lobe

Am

ygal

e

dys

pla

sia

Dai

lyM

27

22

5 R

igh

t

tem

por

al

and

fron

tal l

obes

Hip

poc

amp

us

wit

h

neu

ron

al

loss

and

glio

sis,

con

sist

ent

wit

h

hip

poc

amp

al

scle

rosi

s1–

2/m

onth

M

53

19

34

Bil

ater

al

tem

por

al

lobe

s

Met

asta

tic

mal

ign

ant

mel

anom

a

2–3/

mon

thF

51

3

48

Rig

ht

ante

rior

med

ial t

emp

oral

lobe

Ch

ron

ic

reac

tive

glio

sis

1–3/

mon

thM

26

1.2

16

Tem

por

al

and

med

ial f

ron

tal l

obes

Inte

nse

reac

tive

fibr

illa

ry

astr

ocyt

osis

1–3/

day

X. Zhang et al. / Journal of Neuro

. Methods

.1. Data acquisition

fMRI data of 52 control subjects and 16 patients who sufferedrom intractable medial temporal lobe epilepsy were imaged on aT Siemens Trio scanner at the Yale MRRC.

Table 1 shows the clinical data of the 16 patients (7 malend 9 female) with ages from 11 to 53. History of seizure rangesrom 3 to 48.5 years. All subjects gave informed written con-ent and this study was approved by the Yale IRB. A T1-weighted-plane localizer was used to localize the slices to be obtainednd T1 anatomic scans were collected in the axial-oblique ori-ntation parallel to the ac-pc line. Resting state functional dataas obtained using a gradient echo T2*-weighted echo planar

maging sequence with TR = 1550 ms, TE = 30 ms, flip angle = 80,OV = 22 × 22 cm, matrix size 64 × 64, 25 slices, skip 0 mm, func-ional voxel size 3.4 mm × 3.4 mm × 6 mm. 3–8 runs of resting stateata were collected in interleaved acquisition mode with 229 vol-mes per run.

.2. Data preprocessing

The functional data preprocessing pipeline included the fol-owing steps: first, slice timing correction followed by motionorrection was performed using SPM5; the linear trends in theimecourses were removed; the data were then temporally filteredsing a band-pass filter (0.01–0.1 Hz); and then spatially smoothedith a Gaussian filter of FWHM = 8 mm; finally, the six rigid-bodyovement confounds were removed (Friston et al., 1996).

.3. ROI definition

To investigate a wide range of possible nodes in the lim-ic/medial temporal regions, 36 ROIs were defined in MNI space as

isted in Table 2. This set of ROIs was composed of both anatomicallynd functionally defined regions. The first 28 ROIs were adoptedrom an online map of Brodmann’s areas (BioimageSuite.org). Theast 8 ROIs were functionally defined from BOLD task activationaradigms aimed at language and motor systems (Arora et al.,009).

One approach to performing a network analysis is to map eachndividual subject’s fMRI data into a standard space such as MNIpace. However, the disadvantage of this method is that unex-ected global and local confounds could be introduced into theimecourses in the transformation space due to the interpolationy registration algorithm (Grootoonk et al., 2000). Instead, in thisork the ROI atlas was mapped from MNI common reference space

nto each individual subject’s space. For each ROI, the mean time-ourse within its region was calculated and used to measure theonnectivity matrix which was then input to the network analysis.

.4. Connectivity calculation

A symmetric connectivity matrix A was obtained by correlatinghe mean timecourses of 36 predefined ROIs for each subject. Theimecourses of all scan runs of each ROI were combined into oneector in connectivity matrix calculation.⎛

⎜ a1,1, a1,2, . . . , a1,NV

⎞⎟

= ⎝ a2,1, a2,2, . . . , a2,NV

aNV,1, aNV,2, . . . , aNV,NV

⎠ (1)

= |A| (2) Tab

le

1C

lin

ical

dat

a.

Pati

ent

Ind

1 2 3 4 5 6 7 8 9 10

11

12

13

14

15 16

132 X. Zhang et al. / Journal of Neuroscienc

Table 2ROI list.

ROI index ROI name

1 Right Caudate2 Left Caudate3 Right Putamen4 Left Putamen5 Right Thalamus6 Left Thalamus7 Right Globus Palidus8 Left Globus Palidus9 Right Amygdala10 Left Amygdala11 Right Hippocampus12 Left Hippocampus13 Right BA32 (Superior Anterior Cingulate)14 Left BA32 (Superior Anterior Cingulate)15 Right BA24 (Inferior Anterior Cingulate)16 Left BA24 (Inferior Anterior Cingulate)17 Right BA23 (Inferior Posterior Cingulate)18 Left BA23 (Inferior Posterior Cingulate)19 Right BA31 (Superior Posterior Cingulate)20 Left BA31 (Superior Posterior Cingulate)21 Right Lateral OFC (Orbitofrontal Cortex)22 Left Lateral OFC (Orbitofrontal Cortex)23 Right Medial OFC (Orbitofrontal Cortex)24 Left Medial OFC (Orbitofrontal Cortex)25 Right Anterior Insula26 Left Anterior Insula27 Right Posterior Insula28 Left Posterior Insula29 Right Stg (Superior Temporal Gyrus)30 Left Stg (Superior Temporal Gyrus)31 Right Broca32 Left Broca33 Right Motor

C

cmm

C

newc

D

So

ftrmt

34 Left Motor35 Right Pmtg (Posterior Medial Temporal Gyrus)36 Left Pmtg (Posterior Medial Temporal Gyrus)

where ai,j is the correlation coefficient of ROI i and j.

= B(

VG − NV

VB − NV

)(3)

B is the absolute matrix of A. Since there is a large variation of theonnectivity matrix across subjects, it is necessary to normalize theatrix before it can be compared across subjects. Here, the absoluteatrices B were then normalized based on the sum of matrix:

= B(

VG − NV

VB − NV

)

where VB =∑NV

i=1

∑NVj=1bi,j , VG = W/NC

∑NCk=1(VB)k, NV is the

umber of ROIs, NC is the number of control subjects, bi,j is thentry of matrix B, VG is the weighted group mean, the weight Was selected as 0.85 in this study to ensure most of the metric

omponents of C are lower than 1. VG is 588 given the above W.Finally, all the elements below a fixed threshold were set to 0.

NV×NV ,

⎧⎪⎨⎪⎩

di,j = 1 if i = j

di,j = ci,j if i /= j, ci,j ≥ Th

di,j = 0 if i /= j, ci,j ≥ Th

(4)

The selection of the threshold Th in Eq. (4) will be discussed inections 3.5.1.1 and 3.5.2 where the optimal threshold Th = 0.4 wasbtained based on the optimal classification result.

Fig. 1(a) and (b) is the visualization examples of a network Dor a normal healthy volunteer and a patient respectively where

he network threshold was set to a correlation of 0.4. Each nodeepresents one ROI while the weight between 2 nodes is the nor-alized connectivity value in D. The distance between 2 nodes is

he inverse of the weight. The layout of Fig. 1 was created using

e Methods 199 (2011) 129– 139

the Kamada–Kawai algorithm (Kamada and Kawai, 1989) which isa force based layout method. The nodes located in the center of thegraph have more connections with other nodes while the nodes atthe periphery of the graph have fewer connections.

3.5. Classification

In Section 3.5.1 below, the major concepts and methods used inclassifier construction are discussed. Section 3.5.2 shows the crossvalidation of the proposed classifier in the given data set.

3.5.1. Classifier construction3.5.1.1. Setting a threshold for connectivity matrix. The first step inanalyzing networks using resting-state connectivity data is to set athreshold Th to Eq. (4). The range of the Th is [0,1]. The optimal Thwill be obtained in Section 3.5.2 based on the best classification per-formance. This threshold really represents the correlation betweenthe time-courses for a pair of ROIs with 1 representing completecorrelation and 0 reflecting no correlation in the time-courses.Setting a threshold implies that two regions will be considered con-nected if the correlation coefficient between them is higher than thechosen threshold and they will not be considered connected if thecorrelation coefficient is below this threshold.

3.5.1.2. Boxplot and feature definition. In order to evaluate the pos-sible abnormal network metrics and abnormal ROIs, a Boxplotanalysis was applied.

Boxplots were created for each of 36 ROIs using the 5 socialnetwork measures and the control group data. Let Pij be the jthnetwork property of ith ROI.

Oi,j ={

1 if Pij is outlier

0 otherwisei = 1, 2, . . . , 36, j = 1, 2, . . . , 5 (5)

Each of the 5 network properties across 36 ROIs were consideredas a feature and entered into a feature matrix O.

3.5.1.3. Classification criteria and feature selection. At the individualsubject level, the outliers indicated in matrix O of Eq. (5) reveal localabnormalities in the functional network. Different MTLE patientshave different ROI areas labeled as outliers likely due to individ-ual differences in the MTLE. However, our goal was to investigatewhether or not there were common network properties or abnor-mal ROIs shared within this MTLE patient group. To investigatethe abnormal network properties at the group level, a classifica-tion analysis was performed to determine if the outliers for eachsubject in all the 36 ROIs and 5 network measures could separatethe patient from the control group.

J =36∑i=1

5∑j=1

Oi,j (6)

A suitable classification threshold can be selected for J based onthe trade-off between sensitivity and specificity to separate the twoclasses. Here, the classification threshold was selected by maximiz-ing the following term:

argmax{Sensitivity + Specificity} (7)

where sensitivity and specificity are defined as below:

Sensitivity = true positives

true positives + false negatives× 100% (8)

Specificity = true negatives

ture negatives + false positives× 100% (9)

Patients correctly classified represent true positives, while cor-rectly classified control subjects are true negatives. The patients

X. Zhang et al. / Journal of Neuroscience Methods 199 (2011) 129– 139 133

Fig. 1. Network examples of patient and healthy control subject created using Kamada–Kawai algorithm. (a) Normal subject. (b) MTLE patient. The nodes located in thecenter of the graph have more dense connections with other nodes while the nodes at the periphery of the graph have fewer connections. It can be seen in (b) that the nodes2 rbitoT locatet

tfs

stw‘sh‘

J

w

Note that feature selection was conducted in the training set.

(Left Caudate), 7 (Right Globus Palidus), 12 (Left Hippocampus), 23 (Right Medial Ohese nodes were highlighted with circles in (b). The difference between the nodes

han healthy control.

hat were incorrectly classified as control subjects were labeledalse negatives, and the control subjects that were incorrectly clas-ified as patients were considered false positives.

Classification criteria can be calculated within a feature sub-et and since the sample size in this study is small relativeo the high dimensional feature space, traditional recursive for-ard/back forward feature selection strategy is vulnerable to

over-fitting’. Therefore, we applied the following feature selectiontrategy to select a meaningful feature subset in order to achieveigher classification accuracy while avoiding the problem of

over-fitting’.

=∑

O (10)

(i,j) ∈ FS

i,j

here FS is the meaningful feature subset.

frontal Cortex), 36 (Left Posterior Medial Temporal Gyrus) have fewer connections.d at the periphery of the graph and the other nodes is more distinct in MTLE patient

For each feature, the outlier rate ORA is defined as follows:

ORA = Number of outlier subjects

Total Number of subjects in the group(11)

Feature subsets can be selected based on the outlier rates in bothgroups. Only those features that have a greater outlier rate in thepatient group than control group are selected.

FS = {(i, j)|ORAPatienti,j > ORAControl

i,j }

i = 1, 2, . . . , 36, j = 1, 2, . . . , 5 (12)

Compared with traditional recursive forward/backward featureselection strategy, this approach can avoid ‘over-fitting’ and thusachieve good generalization ability in a small sample set wherehigh variation exists in both groups.

134 X. Zhang et al. / Journal of Neuroscienc

Table 352-Fold cross validation (leave one out).

LOOP 1: Iteratively select 1 out of 52 control subjects as testing sample(TESTING − Control), other 51 control subjects were used as training set(TRAINING − Control) for LOOP 2.

LOOP 2: Iteratively select 1 out of 16 patients as testing sample(TESTING − Patient). Use other 15 patients as patients’ sample of training set(TRAINING − Patient). Select those features that have larger outlier rate incontrol subjects (TRAINING − Control) than in patients (TRAINING − Patient)as the feature subset and calculate the classification threshold via Eq. (7).Test this feature subset and classification threshold in the testing set(TESTING − Control + TESTING − Patient).

End LOOP2End LOOP1Feature subset selection: Iteratively select 1 out of 51 subjects’ training set

3ca1tRwToctA

3

ssctsva(tc

s

(

((

ao

4

4

DBfiltl

(TRAINING − Control) in LOOP 2 together with 15 patients as validationsample, other 50 control subjects were used to define outlier of boxplot.

.5.1.4. Evaluation of classifier performance. The standard tool forontrolling the trade-off between sensitivity and specificity of anlgorithm is a Receiver Operating Characteristic (ROC) curve (Metz,986). The ROC curve is a plot of sensitivity versus 1-specificity (orrue positive rate versus false positive rate). The area under theOC curve (AUR) summarizes the quality of classification over aide range of misclassification costs (Hanley and McNeil, 1983).

he greater the area under the ROC curves the higher the probabilityf making a correct decision. Therefore, AUR was chosen as theriteria of evaluating the performance of classifier in this study. Inhe case of iterative feature selection, Eq. (12) is actually equal toUR > 0.5.

.5.2. Cross validationFor cross validation the training samples were divided into k

ubsets, each of which had the same number of samples. The clas-ifier was then trained k-times: In the ith (i = 1, . . ., k) iteration, thelassifier is trained on all subsets except the ith onset, for whichhe classification error is computed. When k is equal to the totalample number N, this approach is also called ‘leave one out’ crossalidation. It is known that the average of error calculated in thebove loop is a rather good estimate of the generalization errorMartin and Hirschberg, 1996). In order to evaluate the classifica-ion performance, a 52-fold cross validation (leave one out) whichontained 2 loops was conducted as outlined in Table 3:

The overall procedure of the optimal connectivity thresholdelection can therefore be summarized as follows:

a) Setting a connectivity matrix threshold Th with a range of [0,1]in Eq. (4).

b) Running 52-fold cross validation using leave-one-out.c) Calculating AUR of the classifier in step (b).

Performing the steps (a)–(c) iteratively, the optimal Th = 0.4ssociated with the largest AUR can be obtained as shown in Fig. 4f Section 4.4.

. Results

.1. Boxplot of five network metrics in ROIs

Fig. 2 shows boxplot examples for the network measures of (a)egree, (b) Strength, (c) Closeness, (d) Clustering coefficient and (e)etweenness centrality using all the 52 control subjects. The datarom the 16 patients are represented by the circles. Each column

n the figure represents a boxplot for a different ROI. The markersocated above the upper whisker represent positive outliers whilehe markers below the lower whisker represent the negative out-iers. The outliers from each patient can be examined using the

e Methods 199 (2011) 129– 139

boxplot but this work is focused on the analysis of the abnormal net-work properties and regions at the group level. For example, 10/16patients were labeled below the whisker of 11th ROI in Fig. 2(b). Inother words there were 10 patient outliers for the network propertyStrength in the Right Hippocampus ROI.

Different network thresholds emphasize different outlier direc-tions for some network measures and it is useful to examine theeffect of threshold on the outliers. For example, with the networkthreshold 0.4, the 20th ROI, Left BA31 (Superior Posterior Cingulate)has an upper whisker near the upper bound 35 for network mea-sure Degree as shown in Fig. 2(a). Therefore, there are no outliersfor this feature although the data distribution for the patient groupis different from that of the control group in this feature. However,when the network threshold is selected as 0.5, the upper whiskerof this ROI is decreased with only 8 patients labeled as positive out-liers. On the contrary, when the network threshold was increasedfrom 0.4 to 0.5, the lower whisker of the network measure Strengthof 12th ROI, the Right Hippocampus, was decreased to close thelower bound, the negative outlier rate of patient group was reducedfor this feature. The threshold effect is discussed further in Section4.4.

4.2. Feature rank

Table 4 shows the feature rank list based on the area under theROC curve for each of the features. The outlier rates of patients andcontrol subjects were obtained from the boxplots in Fig. 2. Therewere a total of 93 features with AUR > 0.5. Only the top 30 featuresare listed in Table 4. With the network threshold 0.4, the most sig-nificant features are Strength and Closeness of Right Hippocampus.The AUR of both of these features was 0.7163, with the outlier rateof patient group equal to 0.625, and the outlier rate of control groupequal to 0.1963. When the network threshold was changed from 0.4to 0.5, some specific features had quite different rankings, but therewas more than 60% overlap between ROIs that appeared in the top30 features in both cases. ROIs such as Left/Right Hippocampus,Left/Right Caudate, Left/Right BA31 (Superior Posterior Cingulate),and Left/Right BA23 (Inferior Posterior Cingulate) were within thetop 30 features at both thresholds.

4.3. Comparison of classification performance with and withoutfeature selection

Note that the feature selection was conducted in the trainingdata set of each iteration rather than the whole data set in thecross validation. Only those features that had larger outlier ratesin the patient group relative to the control group, in other words,AUR > 0.5, were selected as a feature subset in the classification.These features contribute to the final classification accuracy and assuch were considered meaningful features for the classifier. Includ-ing features with AUR < 0.5 would reduce the classification accuracyas they do not provide information that assists in separating the twogroups and therefore should be excluded.

Fig. 3 demonstrates the comparison of classification per-formance with and without feature selection in 52-fold crossvalidation. The AURs were 0.8838 and 0.8132 respectively. It canbe seen that the feature selection strategy based on Eq. (12) canimprove the classification accuracy.

4.4. Comparison of classification performance with differentthresholds

Fig. 4 shows a comparison of classification performance with 10different connectivity matrix thresholds using 52-fold cross vali-dation. Feature selection was included in all 10 cases. The AURswere 0.8838 (Th = 0.4), 0.8338 (Th = 0.425), 0.8299 (Th = 0.5), 0.8184

X. Zhang et al. / Journal of Neuroscience Methods 199 (2011) 129– 139 135

a

b

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 360

5

10

15

20

25

30

35

Deg

ree

ROI

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36

0

2

4

6

8

10

12

14

16

18

20

Stre

ngth

ROI

c

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Clo

sene

ss

ROI

Fig. 2. Boxplot examples of Degree, Strength, Closeness, Clustering coefficient and Betweenness centrality. Boxplots were created by 52 control subjects. The marks represent16 patients’ data. Note some marks were overlapped.

136 X. Zhang et al. / Journal of Neuroscience Methods 199 (2011) 129– 139

e

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36-0.1

-0.05

0

0.05

0.1

0.15

Bet

wee

nnes

s C

entra

lity

ROI

d

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Clu

ster

ing

Coe

ffici

ent

ROI

Fig. 2. (Continued ).

Fig. 3. Comparison of ROC curves obtained by feature selection and without featureselection. The area sizes under ROC curve are 0.8838 and 0.8132 respectively.

Fig. 4. Comparison of ROC curves with different connectivity matrix thresholds,feature selection was applied in all 10 cases. The largest AUR = 0.8838 was achievedwhen Th = 0.4.

X. Zhang et al. / Journal of Neuroscience Methods 199 (2011) 129– 139 137

Table 4Feature rank based on the area under ROC (AUR), the top 30 features.

Feature rank ROI name Network property AUR Outlier rate(patient group)

Outlier rate(control group)

1 Right Hippocampus Strength 0.716346 0.625000 0.1923082 Right Hippocampus Closeness 0.716346 0.625000 0.1923083 Right Lateral Orbital Frontal Cortex (OFC) Clustering Coefficient 0.700120 0.438702 0.0384624 Right Caudate Clustering Coefficient 0.681490 0.439904 0.0769235 Right Caudate Closeness 0.680889 0.438702 0.0769236 Left Globus Palidus Closeness 0.674279 0.406250 0.0576927 BA32 (Superior Anterior Cingulate) Clustering Coefficient 0.670673 0.437500 0.0961548 Left BA31 (Superior Posterior Cingulate) Strength 0.668870 0.376202 0.0384629 Left Caudate Strength 0.668269 0.375000 0.03846210 Left Caudate Clustering Coefficient 0.659255 0.376202 0.05769211 Left Globus Palidus Strength 0.647236 0.409856 0.11538512 Left BA23 (Inferior Posterior Cingulate) Strength 0.637019 0.312500 0.03846213 Right Hippocampus Clustering Coefficient 0.629808 0.375000 0.11538514 Left Hippocampus Degree 0.627404 0.312500 0.05769215 Right BA32 (Superior Anterior Cingulate) Clustering Coefficient 0.625000 0.346154 0.09615416 Right pmtg Clustering Coefficient 0.617788 0.312500 0.07692317 Left Globus Palidus Degree 0.615986 0.251202 0.01923118 Right motor Betweenness Centrality 0.608173 0.312500 0.09615419 Left Caudate Closeness 0.606370 0.251202 0.03846220 Right Globus Palidus Closeness 0.606370 0.251202 0.03846221 Left BA31 (Superior Posterior Cingulate) Closeness 0.605769 0.250000 0.03846222 Left Lateral OFC Clustering Coefficient 0.605769 0.250000 0.03846223 Right Caudate Betweenness Centrality 0.598558 0.312500 0.11538524 Left Putamen Betweenness Centrality 0.588942 0.312500 0.13461525 Right Medial OFC Closeness 0.588341 0.253606 0.07692326 Left Hippocampus Clustering Coefficient 0.587139 0.251202 0.07692327 Right Globus Palidus Degree 0.584135 0.187500 0.01923128 Right BA23 (Inferior Posterior Cingulate) Closeness 0.575721 0.189904 0.038462

hss

(0tobtsab

5

T(r

imaaecpiiatrilatTf

29 Right BA31 (Superior Posterior Cingulate) Strengt30 Right BA31 (Superior Posterior Cingulate) Closene

Th = 0.375), 0.7922 (Th = 0.475), 0.766 (Th = 0.35), 0.7536 (Th = 0.3),.7501 (Th = 0.45), 0.7285 (Th = 0.525) and 0.588 (Th = 0.55) respec-ively. Since different network thresholds may highlight differentutliers, overall classification accuracies can be different. Theest classification performance was achieved when the networkhreshold was set as 0.4. In this case, by setting an appropriate clas-ification threshold for the total number of outliers in Eq. (10), anverage sensitivity of 77.2% and specificity of 83.86% were achievedy 52-fold cross validation.

. Discussion

The outlier rates of the five network properties summarized inable 4, indicate that ROIs such as the Hippocampus, Caudate, BA31Superior Posterior Cingulate), BA23 (Inferior Posterior Cingulate),epresent the network nodes most affected in MTLE.

The hippocampus is part of the limbic system and plays anmportant role in epilepsy. Hippocampal sclerosis is the most com-

only visible type of tissue damage in temporal lobe epilepsylthough whether the epilepsy is caused by this hippocampalbnormality or the hippocampus is damaged by the cumulativeffects of seizures remains uncertain. Decreased basal functionalonnectivity within epileptogenic networks associated with MTLEreviously has been reported (Bettus et al., 2009) in a study using

nterictal fMRI data. The authors hypothesized that the pathophys-ological alterations such as metabolic and hemodynamic changesssociated with epilepsy may affect the neurovascular couplinghereby altering the relationship between BOLD signal and neu-onal activity. Since it is well known that the hippocampus isnvolved in MTLE it is not unexpected that we observe high out-ier rates on the network measures of Strength, Closeness, Degree

nd Clustering coefficient in hippocampus but this finding supportshe utility of measuring these variables in resting-state fMRI data.he smaller Degree measure in patients suggests that there areewer strong connections (greater than threshold) between the hip-

0.575721 0.189904 0.0384620.575721 0.189904 0.038462

pocampus and the other ROIs in MTLE. The smaller Strength valuesalso indicate the functional connection of the hippocampus withthe other ROIs appears to be generally weaker in MTLE patientscompared with normal controls. The decrease observed in patientsfor the Closeness measure in the hippocampus also arises fromfewer and weaker connections – and therefore increased distance.Outliers were found in the hippocampus in 12 out of 16 patients.More specifically, for those 10 patients whose hippocampus wasclearly labeled as abnormal (Table 1), outliers were found in 9 ofthe 10 patients. It should be noted that the abnormal seizure regionslabeled in Table 1 were detected by icEEG obtained in the ictalstate while the outlier regions detected here were obtained fromresting-state fMRI data collected in the interictal state.

The caudate nucleus is known to be an important part of thebrain’s learning and memory system and as part of this systemit needs to communicate with the hippocampus. Measurementsof grey matter content using MRI and voxel based morphometryhave revealed significant decreases in caudate volume in patientswith refractory medial temporal lobe epilepsy (Bonilha et al., 2004).In this work, the caudate demonstrates high outlier rates in thenetwork measures of Strength, Closeness, Clustering coefficient andBetweenness centrality, a finding that also supports the use of thesemeasures in MTLE. Most of the outliers in Strength and Closenessof caudate are negative outliers. However, there are almost equalnumbers of outliers in both the negative and positive directions forthe Clustering coefficient suggesting a high variation for this featurein the MTLE group compared with that of the healthy controls.

The third major abnormal node revealed by this network analy-sis is the Cingulate gyrus which also functions as an integral part ofthe limbic system, and is involved with emotion formation and pro-cessing, learning and memory, as well as the default mode network.

The network measures of Strength, and Closeness for BA31 (SuperiorPosterior Cingulate) and BA23 (Inferior Posterior Cingulate) rank inthe top 30 abnormal features. In the MTLE epilepsy patients thefunctional connections of BA31 and BA23 with other ROIs were

1 scienc

shfrccisa

acstfwmoo

snhch

pnu

prmpi

uttcpMgrgM(fftTclisa

wntllc

um

38 X. Zhang et al. / Journal of Neuro

tronger than those of normal healthy volunteers according to theigh positive outlier rates of Strength, and Closeness. The increase of

unctional connectivity in Posterior Cingulate Cortex has also beeneported in other studies of MTLE (Zhang et al., 2010). This increasedonnectivity in resting-state has been suggested to be involved inompensatory mechanisms (Bettus et al., 2009). Another possiblenterpretation is that the PCC is involved in initiation of spike andlow-wave discharges activity (Wang et al., 2011) and hence showsn up-regulated network in these patients.

The globus pallidus is a sub-cortical structure of the brain and is major component of the basal ganglia system. It has been impli-ated functionally to play an active part in pre-filtering externaltimuli and may help reduce the amount of irrelevant informa-ion the brain needs to store. In terms of epilepsy there is evidenceor a role of the globus pallidus in the hippocampal epilepsy net-ork (Sabatino et al., 1984; Sawamura et al., 2002). The networkeasures of Degree, Strength and Closeness showed high negative

utlier rates suggesting weak connections between this ROI andther parts of the MTLE network.

Finally, the orbitofrontal cortex (OFC), also part of the limbicystem, has previously also been found to be involved in the MTLEetwork (Blumenfeld et al., 2004). Our network analysis indicatedigh positive outlier rates for the network measure of Clusteringoefficient particularly in the left lateral OFC, which represents theigh local embeddedness of the left OFC in the MTLE network.

Taken together, these findings of alterations in the networkroperties in the limbic system in MTLE patients suggest that suchetwork measures may be valuable in assessing MTLE patientssing resting-state fMRI data.

By selecting those features that have larger outlier rates in theatient group relative to healthy volunteers, a classification accu-acy of AUR = 0.8838 was achieved. These findings reveal that theedial temporal lobe/limbic system exhibits abnormal network

roperties at the group level in fMRI data obtained during the inter-ctal state.

There are other literatures on the studying of interictal MTLEsing fMRI. Liao et al. (2010) reported altered small world proper-ies of the interictal MTLE network with fMRI. A number of areas inhe default mode network including bilateral inferior frontal oper-ular gyrus, left posterior cingulated gyrus, left precunes and rightrecentral gyrus, were shown significant decreases in Degree inTLE. Based on n-to-1 connectivity, some ROIs such as bilateral

yrus rectus, left superior frontal gyrus, left middle temporal gyrus,ight middle orbitofrontal gyrus and right superior medial frontalyrus were shown to have significantly increased connectivity inTLE. Zhang et al. (2009) applied independent component analysis

ICA) in a resting-state fMRI study of temporal lobe epilepsy. Theunctional connectivity of the dorsal attention network (DAN) wasound to be decreased in patients. Morgan et al. (2010) used 2dTCAo investigate the activation of functional network in left MTLE.he connectivity in the left anterior hippocampus, bilateral insularortex and default-mode network was investigated and using theeft anterior hippocampus as a seed region, the authors reportedncreased negative connectivity in the patients compared to controlubjects across a network including thalamic, brainstem, frontalnd parietal regions.

In our study, unlike the binary matrix used in Liao et al. (2010), aeighted undirected connectivity matrix was adopted and matrixormalization was proposed. Using five local network proper-ies, abnormal network properties were found in medial temporalobe/limbic system including hippocampus, cingulated gyrus andateral OFC as well as the caudate and globus pallidus based on the

lassification efficiency.

The nodes of the network analysis in this study were selectedsing predefined ROIs. Connectivity analysis is a region basedethod that relies heavily on the choice of the ROIs used for the

e Methods 199 (2011) 129– 139

analysis. Generalizing the ROI approach used here and performingnetwork analyses on a voxel basis has recently been shown to revealdifferences between patients and control subjects in Alzheimer’sdisease (Buckner et al., 2009). Such an approach can reveal localtopology of hubs revealing both ‘connector hubs’ (hub regions thatare highly connected within one module) and ‘provincial hubs’ (hubregions that link different modules) (Sporns et al., 2007; Hagmannet al., 2008) in epilepsy networks. Work is underway in our lab(Shen et al., 2010) to develop cortical parcellation approaches basedon graph theory that can use resting-state data to define functionalsubunits (nodes) for this type of analysis thereby decreasing theneed for reliance on anatomically or atlas defined ROIs.

6. Conclusions and future work

The findings presented here indicate that network propertiesmeasured from resting-state fMRI data collected in the interictalstate may reveal abnormal nodes involved in the epileptogenic net-work. Network theory provides measures such as Degree, Strength,Closeness, Clustering coefficient and Betweenness centrality, thatappear to serve as efficient features that can distinguish healthyvolunteers from medial temporal lobe epilepsy patients, even whenthese measures are based on data collected in the interictal state.They can extract meaningful local information about networkabnormalities associated with epilepsy and identify specific nodeswith abnormal connections. High outlier rates for some of the net-work metrics were found in the limbic system with the primaryregions including the hippocampus, cingulated gyrus and lateralOFC together with caudate and globus pallidus providing the mostdiscerning information. These findings support the notion that net-work analysis can reveal altered patterns of connectivity changesthat are consistent across patients in a clinical category.

In this work we have introduced a functional connectivitymatrix normalization method and a feature selection strategydesigned to improve the classification performance of these net-work properties. Based on the classification criteria, an averagesensitivity of 77.2% and specificity of 83.86% was achieved using52-fold cross validation.

Further investigations on methods to improve the classifica-tion accuracy, are underway using other measurements such ascoherence and partial correlation/coherence as the input to theconnectivity matrix.

Disclosure of conflict of interest

None of the authors have any conflicts of interest to disclose.

Acknowledgment

NIH R01 EB009666-01A1.

References

Arora J, Pugh K, Westerveld M, Spencer S, Spencer DD, Constable RT. Languagelateralization in epilepsy patients: fMRI validated with the Wada procedure.Epilepsia 2009;50:2225–41.

Basset DS, Bullmore E. Small-world brain networks. Neuroscientist 2006;12:512–23.Bertashius KM. Propagation of human complex-partial seizures: a correlation anal-

ysis. Electroencephalogr Clin Neurophysiol 1991;78:333–40.Bettus G, Guedj E, Joyeux F, Confort-Gouny S, Soulier E, et al. Decreased basal fMRI

functional connectivity in epileptogenic networks and contralateral compen-satory mechanisms. Hum Brain Mapp 2009;30:1580–91.

Biswal B, Yetkin FZ, Haughton VM, Hyde JS. Functional connectivity in the motor

cortex of resting human brain using echo-planar mri. Magn Reson Med1995;34:537–41.

Bonilha L, Rorden C, Castellano G, Pereira P, Rio PA, Cendes F, et al. Voxel-based mor-phometry reveals gray matter network atrophy in refractory medial temporallobe epilepsy. Arch Neurol 2004;61:1379–84.

scienc

B

B

D

F

G

H

H

H

H

H

H

K

K

L

L

M

MM

M

work in temporal lobe epilepsy: a resting FMRI study. Neurosci Lett 2009;458:97–101.

X. Zhang et al. / Journal of Neuro

uckner RL, Sepulcre J, Talukdar T, Krienen FM, Liu H, Hedden T, et al. Cortical hubsrevealed by intrinsic functional connectivity: mapping, assessment of stability,and relation to Alzheimer’s disease. J Neurosci 2009;29:1860–73.

lumenfeld H, Rivera M, McNally KA, Davis K, Spencer DD, Spencer SS. Ictal neocor-tical slowing in temporal lobe epilepsy. Neurology 2004;63:1015–21.

ellen EV, Douw L, Baayen JC, Heimans JJ, Ponten SC, Vandertop WP, et al. Long-term effects of temporal lobe epilepsy on local neural networks: a graphtheoretical analysis of corticography recordings. PLoS One 2009;4(11):e8081,doi:10.1371/journal.pone.0008081.

riston KJ, Williams S, Howard R, Frackowiak RS, Turner R. Movement-related effectsin fMRI time-series. Magn Reson Med 1996;35:346–55.

rootoonk S, Hutton C, Ashburner J, Howseman AM, Josephs O, Rees G, et al. Char-acterization and correction of interpolation effects in the realignment of fMRItime series. Neuroimage 2000;11:49–57.

agmann P, Cammoun L, Gigandet X, Meuli R, Honey CJ, Wedeen VJ, et al. Mappingthe structure core of human cerebal cortex. PLoS Biol 2008;6(7):e159.

ampson M, Peterson BS, Skudlarski P, Gatenby JC, Gore JC. Detection of func-tional connectivity using temporal correlations in MR images. Hum Brain Mapp2002;15:247–62.

ampson M, Driesen N, Skudlarski P, Gore JC, Constable RT. Brain connectivityrelated to working memory performance. J Neurosci 2006a;26(51):13338–43.

ampson M, Tokoglu F, Sun Z, Schafer RJ, Skudlarski P, Gore JC, et al. Connectivity-behavior analysis reveals that functional connectivity between left BA39 andBroca’s area varies with reading abiliy. Neuroimage 2006b;31(2):513–9.

anley J, McNeil B. A method of comparing the areas under receiver operatingcharacteristic curves derived from the same cases. Radiology 1983;148:839–43.

orstmann MT, Bialonski S, Noennig N, Mai H, Prusseit J, Wellmer J,et al. State dependent properties of epileptic brain networks: comparativegraph–theoretical analyses of simultaneously recorded EEG and MEG. Clin Neu-rophysiol 2010;121:172–85.

amada T, Kawai S. An algorithm for drawing general undirected graphs. InformProcess Lett 1989;31:7–15.

ramer MA, Kolaczyk ED, Kirsch HE. Emergent network topology at seizure onset inhumans. Epilepsy Res 2008;79:173–86.

iao W, Zhang Z, Pan Z, Mantini D, Ding J, et al. Altered functional connectivityand small-world in mesial temporal lobe epilepsy. PLoS One 2010;5(1):e8525,doi:10.1371/journal.pone.0008525.

owe MJ, Mock BJ, Sorenson JA. Functional connectivity in single and multislice echo-planar imaging using resting-state fluctuations. Neuroimage 1998;7:119–32.

artin JK, Hirschberg DS. Small samples statistics for classification error rates I:Error rate measurement. Department of Information and Computer Science, UCIrvine, Tech. Report; 1996.

etz C. ROC methodology in radiological imaging. Invest Radiol 1986;21:720–33.organ VL, Gore JC, Abou-Khalil B. Functional epileptic network in left mesial

temporal lobe epilepsy detected using resting fMRI. Epilepsy Res 2010;88:168–78.

ullen KM, Vohr BR, Katz KH, Schneider KC, Lacadie C, Hampson M, et al. Pretermbirth results in alteration in neural connectivity at age 16 years. Neuroimage2010.

e Methods 199 (2011) 129– 139 139

Nooy WD, Mrvar A, Batagelj V. Exploratory social network analysis with Pajek. Cam-bridge University Press; 2005.

Onnela JP, Saramaki J, Kertesz J, Kaski K. Intensity and coherence of motifs inweighted complex networks. Phys Rev E 2005;71:065103.

Ponten SC, Bartolomei F, Stam CJ. Small-world networks and epilepsy: graph theo-retical analysis of intracerebrally recorded mesial temporal lobe seizures. ClinNeurophysiol 2007;118:918–27.

Reijneveld JC, Ponten SC, Berendse HW, Stam CJ. The application of graphtheoretical analysis to complex networks in the brain. Clin Neurophysiol2007;118:2317–31.

Sabatino M, La Grutta V, Gravante G, La Grutta G. Interrelations between globuspallidus and hippocampal epilepsy in the cat. Arch Int Physiol Biochim1984;92:291–6.

Shen X, Papademetris X, Constable RT. Graph-theory based parcellation offunctional subunits in the brain from resting-state fMRI data. Neuroimage2010;50(3):1027–35.

So NK. Depth electrode studies in mesial temporal epilepsy. Epilepsy surgery. RavenPress; 1991.

Spencer SS, Marks D, Katz A, Kim J, Spencer DD. Anatomic correlates of interhip-pocampal seizure propagation time. Epilepsia 1992;33:862–73.

Spencer SS, Spencer DD. Entorhinal–hippocampal interactions in medial temporallobe epilepsy. Epilepsia 1994;35:721–7.

Spencer SS. Neural networks in human epilepsy: evidence of and implications fortreatment. Epilepsia 2002;43:219–27.

Sporns O, Chialvo DR, Kaiser M, Hilgetag CC. Organization, development and functionof complex brain networks. Trends Cognit Sci 2004;8:418–25.

Sporns O, Honey CJ, Kotter R. Identification and classification of hubs in brain net-works. PLoS One 2007;2(10):e1049, doi:10.1371/journal.pone.0001049.

Spoms O, Rubinov M, Kotter R. Brain Connectivity Toolbox.http://sites.google.com/a/brain-connectivity-toolbox.net/bct/Home.

Stam CJ, Jones BF, Nolte G, Breakspear M, Scheltens P. Small-world networks andfunctional connectivity in Alzheimer’s disease. Cereb Cortex 2007;17:92–9.

Sawamura A, Hashizume K, Tanaka T. Electrophysiological, behavioral and metabolicfeatures of globus pallidus seizures induced by a microinjection of kainic acid inrats. Brain Res 2002;935:1–8.

Wang Z, Lu G, Zhang Z, Zhong Y, Jiao Q, Zhang Z, et al. Brain Res 2011;1374:134–41.Wasserman S, Faust K. SocialNetwork analysis: methods and application. Cambridge

University Press; 1994.Wennberg R, Arruda F, Quesney LF, Olivier A. Preeminence of extrahippocampal

structures in the generation of mesial temporal seizures: evidence from humandepth electrode recordings. Epilepsia 2002;43:716–26.

Zhang Z, Lu G, Zhong Y, Tan Q, Yang Z, Liao W, et al. Impaired attention net-

Zhang Z, Lu G, Zhong Y, Tan Q, Liao W, Wang Z, et al. Altered spontaneous neuronalactivity of the default-mode network in mesial temporal lobe epilepsy. BrainRes 2010;1323:152–60.